VILLASMART: Wireless Sensors For SystemIdentification in Domestic Buildings

Alexandru Caracas, Fabian L. Mueller, Olle Sundstrom, Carl Binding, Bernhard JansenIBM Research - Zurich

CH-8803 Rueschlikon, SwitzerlandEmail: {xan, fmu, osu, cbd, bja}@zurich.ibm.com

Abstract—In this paper, we describe the design decisions, theimplementation caveats, and the challenges faced while deployinga special-purpose wireless sensor network (WSN) at a domesticbuilding that is part of a smart grid test environment. The sensornetwork is designed to collect high-quality data of the building’sin- and outdoor climate, heating system, and energy consumption.This rich sensor data is analyzed in order to characterize thebuilding’s potential to act as a flexible energy consumer in asmart grid framework. In addition to the functional challenges,a continuous, low-power operation over extended periods as wellas a plug&play, non-intrusive installation are required. Moreover,restrictions on the hardware costs and the development-to-deployment time had to be considered. Our simulation-drivenapplication development and our highly modular design provedto be essential in meeting these challenges. This approach andits application to a real-world installation represent the keycontributions of this work.

I. INTRODUCTION

The recent drive to significantly increase the share of re-newable energy sources poses many challenges to the existingelectric grid. One major challenge is the highly fluctuatingproduction of wind and solar power which requires effectivemeans of dynamically balancing the grid in order to compen-sate for peak supply and demand.

In the context of the ECOGRID EU [1] project, our focusis on using the thermal inertia of domestic buildings asenergy buffers with the goal of balancing the power grid.Domestic buildings do not need to constantly consume energyfor heating purposes but can act as flexible energy consumersin the smart grid. However, in order to make the energeticflexibility of domestic buildings predictable and controllable,their energetic behavior needs to be understood and identified.Towards this end, we designed, developed, and installed amodular and extensible wireless sensor network (WSN) in atest and reference household called VILLASMART.

Based on the measurements from the sensor network, wederived and identified a mathematical model describing theenergetic behavior of the building. On the one hand, themodel needs to capture the fundamental thermal dynamicsof the house depending on heat-pump operation and weatherconditions. On the other hand, the model has to incorporatethe behavior of the heat pump and its internal control system.The model is used to quantify and predict the flexibility of thesystem. Moreover, typical smart grid control systems rely onmodel-predictive control schemes when trying to find controlsignals that are optimal with respect to electricity costs or some

objective capturing the grid balancing requirements. The effec-tiveness and performance of such methods fundamentally relyon the availability and accuracy of appropriate mathematicalmodels.

The remainder of this paper is structured as follows. InSec. II, we discuss related work combining WSN and model-based approaches in smart grid contexts. Sec. III introducesthe VILLASMART test facility and the requirements imposedon the WSN. Subsequently, Sec. IV elaborates on our designdecisions considering the given requirements. In Sec. V, wehighlight the learning from the actual development and de-ployment of the WSN. Finally, we describe our approach ofderiving and identifying a thermal building model in Sec. VIbefore concluding in Sec. VII.

II. RELATED WORK

In contrast to industrial building automation, we focus ondomestic buildings, where wireless sensors can be used forretrofitting. In the area of home automation, several wirelesssystems have been developed based on various wireless tech-nologies [2]. More recent efforts [3] include the identificationof thermal models for domestic buildings based on temperaturedata. Our focus is on collecting a richer set of sensor data,e.g. detailed weather and power consumption information, inorder to further improve the accuracy of thermal models. Thisresults in more precise predictions and more efficient control.

A good summary on literature about deploying outdoorsensor networks can be found in [4]. Our work adds tothe existing best-practices by providing insight into indoorWSN deployment in a real home setting. In addition to thelist of “does” and “donts” compiled by [4], we contributewith the following aspects: i) simulation-based applicationdevelopment as a means to decouple and speed-up the softwaredevelopment, ii) hardware integration based on heterogeneous,modular, and extensible components for reducing costs, andiii) built-in management functionality for alleviating deploy-ment efforts.

III. VILLASMART

The VILLASMART home represents the installation envi-ronment for our WSN. It is a fully functional and furnishedbuilding on the Danish island of Bornholm and is used as areference home in the ECOGRID EU project [1]. As shown inFig. 1i, the single-floor building is divided into seven rooms.

i ii iiiFig. 1. The indoor sensor deployment at the VILLASMART reference home. The floor plan (i) shows the positions of the air-temperature sensors (T1-T7). Thedetailed view of the heat-pump room (ii) indicates the positions of the air temperature (T3), water-circuit temperature (W1, W2) and power (C1-C3) sensors.Data streams from each sensor are identified by URIs and are pushed to a messaging broker on a secure channel. Each indoor WSN node (iii) contains amote (a) responsible for both computation and communication, an optional adapter board (b) for different sensors, and a replaceable AA battery pack (c).

An air-water heat pump heats the boiler water as well as thewater of the heating circuit, cf. Fig. 1ii. Photo-voltaic cells onthe building’s roof enable local power generation.

Prior to our WSN deployment, several technologies rangingfrom smart meters, thermostats, and actuators, as well aswireless routers and several servers were already installedat VILLASMART. However, the existing sensor infrastructuredoes not provide the level of detail required for the identifi-cation of our thermal models. However, the existing ICT usesWiFi, Z-Wave, or ZigBee as a means of communication, whichcompete with our WSN for the wireless spectrum.

Our WSN infrastructure at VILLASMART comprises air andwater temperature sensors1, power sensors2, a solar radiationsensor3, as well as an outdoor weather station4. In addition tothe monitoring capabilities, remote actuation allows to control5

the operational state of the heat pump.

IV. DESIGN DECISIONS

To meet the rapid development schedule and a limitedhardware budget, we relied on a heterogeneous set of both off-the-shelf and custom-built components. To enable the paralleldevelopment of hardware and software as well as a modular

1Air and water temperature as well as power consumption exhibit fastdynamics and are, thus, sampled every 10 s.

2The power sensors should allow for a non intrusive installation. For thispurpose, we use split-core current transformers (or clamp readers).

3On sunny days, the amount of solar radiation, when entrapped via largewindows, can significantly influence room temperatures and is thus sampledin 30 s intervals.

4The dynamics of weather data are of minor importance. It suffices tosample in 2 min intervals for outdoor temperature, humidity, barometricpressure, wind speed, wind direction, and rain volume.

5For safety considerations, a local watchdog allows the heat pump to operatebased on its own internal control if no external control was received for acertain time period.

deployment, the WSN was designed as two independent, com-plementary sub-networks satisfying the different requirementsfor indoor and outdoor installation. An embedded PC, locatedin the heat-pump room, cf. Fig. 1ii, serves as the gatewaybetween a secure, broker-based messaging service and thetwo wireless networks. The model analysis and the controllogic interact with the WSN through queues provided by themessaging broker.

A. Note on Wireless Communication

The indoor and outdoor sub-networks have different re-quirements in terms of wireless communication and protocols.However, irrespective of the protocol, the common denom-inator for the physical layer and frame format is the well-established IEEE 802.15.4 standard in the 2.4 GHz6 ISM band.

In a neighborhood, weather data is valid for the individualmodels of several households. Thus, the communication rangeof the outdoor sensors should cover a few hundred meters. Alarger range also means that the outdoor network does notrequire a mesh topology. In this case, a star topology is easierto maintain and is more energy efficient as data rates arerelatively low. Consequently, a carrier sense multiple access(CSMA) network protocol is well suited for this purpose.

By contrast, adjacent homes should be treated individuallyin terms of thermal behavior. Thus, the communication rangeof indoor sensors should cover only a few meters. Walls andmetal enclosures disrupt wireless links. For this reason, a meshtopology is required for covering the entire house. In addition,the sampling rates of the indoor sensors are significantly

6Using the 868 MHz band to escape the interference in the 2.4 GHz bandintroduces additional challenges. A lower frequency offers greater range,which increases the risk of interference between indoor networks frommultiple homes. Furthermore, depending on the encoding, the bandwidth forthe 868 MHz may not suffice for higher sampling rates.

Fig. 2. Architectural overview for the WSN deployment.

higher. We employ an energy-efficient time division multipleaccess (TDMA) network protocol to meet the long life-timerequirements while maintaining the higher sampling rates.Fig. 2 shows an overview of our system architecutre.

B. Choosing Off-the-shelf Hardware

To support a large variety of sensors, cf. Sec. III, it isdesirable to have WSN nodes that easily integrate sensors andadapt to future modifications. As temperature is one of thekey parameters for the thermal models, it is crucial to use thesame sensing element for measuring both air and circuit-watertemperature. For this purpose, digital and factory-calibratedsensors are an appropriate choice. We used a breakout board[5] of the TMP102 sensor which is sufficiently accurate andcan be integrated using the I2C bus.

For the indoor network, we selected the Linear nodes [6]to meet the communication and sensor requirements. Theyprovide direct access to digital and analog GPIO pins, whichfacilitates integration. However, the nodes require customindoor housing, cf. Fig. 1iii, for protection.

In contrast, all the electronic components for the outdoornetwork demand protection (according to the ingress protec-tion standard the appropriate rating is IP67) against dust andwater while being unobtrusive to the sensing elements. Wemainly rely on Waspmote nodes [7] to collect weather data.The nodes provide a long range XBee radio, cf. Sec. IV-A, andappropriate housing for sensors, cf. left-hand side of Fig. 3.

Finally, the solar radiation sensor, cf. right-hand side ofFig. 3, is a low-cost custom design based on readily availablecomponents. The short-circuit current produced by the photo-voltaic panel (d) is analyzed and converted by a correspondingWSN node (e). We relied on a high-end pyranometer forcalibrating the conversion values.

C. Common Functionality for Sensor-specific Applications

In general, each WSN node runs a custom software appli-cation7 tailored for sampling and processing the data fromthe attached sensor(s). Abstracting from the particularities ofeach sensor, a set of common functionality is required acrossall applications for maintenance reasons.

To ease deployment, all applications should provide infor-mation and statistics about the network topology, the quality ofthe wireless signals, and the battery level of the WSN nodes.

7The power sensing application is one example. It collects ADC samplesat 2 kHz for all the three AC phases. The samples are collected for at least80 ms to cover 4 periods of 50 Hz AC. Subsequently, the application computesthe root-mean-square (RMS) of the voltage. The resulting consumed poweris sent as a report every 20 s together with other statistics.

Fig. 3. The outdoor WSN at VILLASMART encompasses two nodes. The left-hand node (a) is installed in the courtyard and connects to the temperature,humidity, and barometric pressure sensors (b) and the wind speed, winddirection, and rain gauge sensors (c). The right-hand node is a custom solarradiation sensor (e) installed on the roof of the house, facing southwards.

The ability to reconfigure parameters at runtime is anotherimportant common functionality. This allows us to selectdifferent sampling rates, thresholds, buffers, and aggregationalgorithms.

Anticipating future requirements, the sensor nodes shouldallow for over-the-air remote application maintenance, e.g.the loading, updating, and deleting of entire applications.Automatic energy management [8], e.g. powering off unusedmodules, is required for extended periods of operation on theinitial set of batteries. In our case, these maintenance andmanagement functions are provided by the Mote Runner [9]OS/VM running on the WSN nodes.

V. DEVELOPMENT AND DEPLOYMENT CHALLENGES

In the following we highlight the challenges faced duringthe development, testing, and the deployment of our WSN.As early on as possible, for a period of three months, wemimicked the actual deployment with a constantly-runningtestbed. The testbed was extended gradually to the full WSNas hardware components became available.

A. Writing The Software Without The Hardware

Due to procurement delays and a tight schedule, not allthe required hardware components were available during thedevelopment phase. Even worse, many of the actual sensingelements were missing. In order to make progress with devel-opment it was vital to simulate the operation including warm-up times, communication, e.g. I2C commands, and the readingof sensor values. The hardware specifications provided thedetails for the script-based simulation within the Mote Runnerenvironment [9].

When compared to other WSN simulations, such as COOJA[10] and Avrora [11], our simulation environment is able toexecute the same application code as is run on the actualhardware. Consequently, no source code modifications wererequired when moving from simulated to real WSN nodes.Moreover, this enabled us to inspect and debug the distributedWSN from within the simulation environment while executingdirectly against back-end systems. Thus, we were able tointegrate the code for the database and message queues earlyin the development process.

B. Providing Protection for the Future Integration

An important aspect of any sensor network is the housing.It protects the electronic components from irreparable damagecaused by transportation, handling, and operation.

At the same time, the housing should be flexible and allowfor integration of different types of sensors. Irrespective ofthe attached sensors, all our indoor nodes use the same fire-repellent and shock-proof housing, cf. Fig. 1iii, which simpli-fies manufacturing. A sufficiently large aperture offers the flex-ibility to connect different sensing elements. For our currentdeployment, we attached up to three current transformers viaTRS audio jacks, and temperature sensors via I2C.

C. Communicating in Spite of Interference

One of the biggest challenges we faced during deploymentwas the wireless communication in the 2.4 GHz band. Asdescribed in Sec. III, the wireless medium is shared with ex-isting infrastructure. To avoid collisions with other equipment,we opted for a channel-hopping, TDMA-based protocol [6]for the indoor WSN. The outdoor WSN is experiencing lessinterference and a CSMA-based protocol provides a viablesolution in this case.

However, the CSMA approach worked well only for theweather station, cf. Fig. 3, which has enough transmissionpower available. In case of the solar sensor, a significantnumber of messages were lost due to a less powerful transmis-sion and the vicinity of other equipment causing interference.Consequently, the solar sensor was reconfigured to act as a leafnode attached to the indoor TDMA network. This could beeasily changed because of our flexible network and hardwaredesign, cf. Sec. IV-B.

D. Anticipating the Extra Steps of Installation

Most of the installation steps, such as connecting currenttransformers and sensors, could be rehearsed in our testbed.In one such rehearsal, we were able to find and fix anintriguing mounting issue concerning the water-pipe sensors.The sensors have to be in physical contact with the pipe foraccurate readings. However, improper mounting to the metalpipes caused sporadic short circuits leading to a node resetor a wrong behavior in the network protocol. An improvedmounting design solved this issue.

Eventually, additional steps were required during the ac-tual installation at VILLASMART. These steps could not berehearsed beforehand. The current transformers required thespecial assistance of an on-site electrician to access the inter-nals of the heat pump. The heat-pump enclosure turned out tobe a Faraday cage and, thus, the corresponding WSN nodessensing the power consumption had to be mounted outsideof it. Additionally, the water pipe sensors are also placed ontop of the heat pump, cf. Fig. 4. Such installation steps requireextra flexibility in terms of both hardware and software, inparticular when frequently attaching and reattaching sensorsand cables, the WSN is expected to continue its operation.

During installation, unexpected situations can and will oc-cur. For example, our outdoor temperature sensor suddenly

Fig. 4. The complexities of an actual installation as reflected by the heat-pumpenclosure. Our WSN nodes (a,b) share space with existing infrastructure.The water-temperature sensors (b) are mounted directly to pipes. The currenttransformers (c) are attached to power lines inside of the enclosure.

failed during deployment. The same sensor had been workingflawlessly on the first installation day. We later found that thefailure was caused by excessive mechanical strain on the cableconnecting the sensing element. This strain occurred wheninserting the sensor into the solar shield. At a later stage,a replacement temperature sensor was shipped and installed.In spite of the sensor malfunctioning, the entire WSN wasoperational at all times, which shows the robustness of themodular approach.

VI. SYSTEM IDENTIFICATION AND EVALUATION

The main purpose of the WSN installed at VILLASMART isto provide the measurement data required for constructing andidentifying a thermal model of the house as well as to giveinsight into the operation of the heat pump. In this section, wepresent initial results on the modeling of the thermal behaviorof the house that was conducted based on sensor data.

A. Thermal Model of VILLASMART

The total thermal energy of VILLASMART, cf. Fig. 1i, isassumed to be stored in only two distinct energy buffers. Theheavy structure of the house comprising walls and floor isdefined to be the first energy buffer and contains the biggestportion of the house’s energy. The second, much smallerenergy buffer represents the energy stored in the building’sheating circuit including the heating water itself, the waterpipes, and the heating devices (radiators).

The behavior of thermal energy buffers as well as cor-responding heat inflows and outflows can be convenientlydescribed by means of a standard resistance-capacitance (RC)model as depicted in Fig. 5. Many other RC networks havebeen proposed to model the thermal dynamics of building,see e.g. [12], [13].

Our model consists of the two energy buffers mentionedabove and represented by the capacitors C1 and C2 in Fig. 5.The corresponding energy levels are given by the tempera-ture of the building’s heavy structure, Tw, and the heatingcircuit temperature, Tc, respectively. Energy can flow fromone buffer to the other only via the house’s internal air,whose temperature Tair is an algebraic function of the two

Fig. 5. The thermal model as an equivalent electric circuit.

buffers’ energy levels. Conductive and convective heat flowresistances, lumped in R1, R2 and R3, govern the exchangeof energy between the house’s solid structure and the in-/outdoor air, as well as between the radiators and the indoorair. For simplicity, the energy supply to the heating circuit isassumed to be directly proportional to the heat pump’s powerconsumption, PHP , where η is the proportionality factor. Theenergy influx due to solar irradiation is approximated by thepower produced by the solar panels, PPV , multiplied by aconstant α. Measurements suggest that the solar energy shoulddirectly affect indoor temperature rather than the house’s heavystructure. The thermal dynamics of the structure are too slowcompared to the indoor air temperature that reacts to solarirradiation almost instantaneously.

Using Kirchoff’s current law, we can derive the differentialequations describing the thermal dynamics:

d

dtTc(t) =

−1

C2(R2 +R3)Tc(t) +

1

C2(R2 +R3)Tw(t)

+η

C2PHP (t) +

αR2

C2(R2 +R3)PPV (t) (1)

d

dtTw(t) =

1

C1(R2 +R3)Tc(t) +

−(R1 +R2 +R3)

C1R1(R2 +R3)Tw(t)

+αR3

C1(R2 +R3)PPV (t) +

1

C1R1Tamb(t) . (2)

By considering [Tc(t), Tw(t)]T as the state,

[PHP (t), Tamb(t), PPV (t)]T as the input, and Tair(t)

as the output of the system, the dynamics (1)-(2) can bewritten as a linear time-invariant state-space system. Themodel’s unknown parameters are the lumped heat flowresistances R1, R2 and R3, the energy buffer’s heat capacitiesC1 and C2, as well as the proportionality constants α and η.

B. Model Identification and Evaluation

Given the model (1)-(2), the goal of the system identificationis to estimate the unknown model parameters based on dataof the system input u(·) provided by the WSN.

We used a gray-box estimation routine of MATLAB’sSystem Identification Toolbox [14] to calculate an estimate ofthe model parameters. Once the values for the parameters areknown and an initial state x(0) has been chosen, the systemcan be simulated and validated.

Fig. 6 compares the measured indoor temperature to ourmodel predictions over a period of approximately 3 days. Themaximum prediction error is 1.79◦C.

VII. CONCLUSION

Indoor and outdoor WSN are essential in future smart homesthat are acting as flexible energy consumers in the power

Fig. 6. Comparison of indoor air temperature measurements and modelpredictions over a representative period of approximately three days.

grid. The sensors provide the required information based onwhich energetic models of domestic buildings can be derived,identified, and evaluated. Accurate models are used to predicta system’s energy consumption and the potential of balancingin the power grid. Moreover, models lie at the core of manypredictive control schemes.

In this paper, we showed that a modular ICT design and asimulation-driven WSN approach offers the required flexibilityand robustness for a successful deployment. Our solution canbe easily integrated in newly-built homes and, alternatively,can be used to retrofit older homes. Furthermore, the practicalexperiences presented in this paper can be extrapolated andserve as a guide for improving future WSN deployments insmart grid contexts.

REFERENCES

[1] EcoGrid EU, http://www.ecogrid-eu.net.[2] L. Liang, et.al., “Design and implementation of wireless smart-home

sensor network based on ZigBee protocol,” in ICCCAS. IEEE, 2008,pp. 434–438.

[3] C. A. Carmody and T. O”Mahony, “System identification of adomestic residence using wireless sensor node data,” in MED. IEEE,2009, pp. 987–992.

[4] G. Barrenetxea, et.al, “The hitchhiker’s guide to successful wirelesssensor network deployments,” in SENSYS. ACM, 2008, pp. 43–56.

[5] Sparkfun, https://www.sparkfun.com/products/9418.[6] Linear, http://www.linear.com/products/smartmesh ip.[7] Libelium, http://www.libelium.com/products/plug-sense/.[8] A. Caracas, et.al., “Energy-efficiency through micro-managing

communication and optimizing sleep,” in SECON. IEEE, 2011, pp.53–63.

[9] Mote Runner, http://www.zurich.ibm.com/moterunner.[10] F. Osterlind, et.al., “Cross-level sensor network simulation with

COOJA,” in LCN. IEEE, 2006, pp. 641–648.[11] B. Titzer, et.al., “Avrora: Scalable sensor network simulation with

precise timing,” in IPSN. IEEE, 2005.[12] J. H. Kampf and D. Robinson, “A simplified thermal model to support

analysis of urban resource flows,” Energy and Buildings, vol. 39,no. 4, pp. 445 – 453, 2007.

[13] Z. O’Neill, et.al., “Model-based thermal load estimation in buildings,”in IBPSA, 2010, pp. 474–481.

[14] “Mathworks MATLAB system identification toolbox,”http://www.mathworks.com/products/sysid/index.html.